The robust unscented Kalman filter (UKF) is revisited in this paper from a new point of view, namely the statistical linear regression (SLR) perspective of the unscented transformation (UT). When the actual distribution of the observation noise deviates from the assumed Gaussian distribution, the performance of the filter may degrade significantly owing to lacking robustness. After linearizing the nonlinear observation equation using the SLR perspective, the M-estimator is employed to replace the weighted least-squares one, resulting in the robust UKF. Besides providing new theoretical aspects, the advantageous performance of the proposed filter is re-emphasized in terms of the following three: (1) robust, the proposed filter is robust against deviations from the assumed Gaussian distribution; (2) efficient, by elaborately selecting the tuning parameter of the M-estimator, the robust filter will have a slight efficiency loss compared with the UKF when the Gaussian assumption does hold; (3) accurate, the proposed filter achieves robustness without compromising the accuracy of the UT.